Hello,

We have a couple of questions regarding the spec of AveragePool and MaxPool when a value in the output tensor is generated by a "pure padding region"

1. AveragePool: input = [[[1, 2, 3, 4, 5]]], kernel_shape = [2], pads = [2, 3], strides = [2], count_include_pad = 0, in this case, the first output value in the output tensor has to be generated by "pure padding region", and since count_include_pad = 0, should the output be nan (division by 0) or 0 (since it did not see any non-padded input) ? I suggest we modify the spec to make the value 0 to reduce ambiguity.

2. MaxPool: For the same example as above, if the first output value is generated by "a pure padding zone", what should the value be? Additionally in the "indices" tensor, what will the first value be? It cannot be in [0, N *C * D1 *... *Dn). I suggest it be a negative value.

What do you guys think ? @houseroad @zrphercule @pranavsharma @tracysh

Thanks!